practice · intuition · agentic engineering — shaped in the open
Build with AI, not just using it.
A living, community-shaped curriculum for the practice of agentic coding.
no install · no signup · free · open in any browser
microsoft.github.io/learn-ai-native-dev |
this repo is the contribution platform
git clone & npm run dev |
For contributors running the site locally. Just learning? Open the tutorial above — nothing to install.
git clone https://github.com/microsoft/learn-ai-native-dev.git
cd learn-ai-native-dev && npm install && npm run devOpen http://localhost:5173. That's it.
React 19 · TypeScript · Vite 7 · Tailwind v4 · shadcn/ui · Radix
This isn't a catalog of skills, agents, or MCP configs. Those are the medium; the product is intuition — knowing which to reach for, when, and with what bounds.
- ☑︎ Tell when a problem wants a one-shot prompt vs. a spec vs. a full agent loop
- ☑︎ Recognize a runaway agent before the token bill does
- ☑︎ Spot the difference between a lazy prompt and a structured one — and feel it in your output
- ☑︎ Move from vibe-prompting → spec-driven → agentic engineering — and know which one a problem calls for
- ☑︎ Pick up MCP, custom agents, and skills as tools you reach for, not lore you memorize
Capability is reference. Intuition is curriculum. —
docs/philosophy.md§2
Agentic coding is barely three years old. Models change every quarter, tools change every month, and yesterday's "best practice" is tomorrow's antipattern. Nobody has the hindsight yet to publish a durable catalog of patterns — not us, not anyone. Anyone who claims otherwise is selling snapshots as truths.
So we're not writing one. We're building the opposite: a place to get reps on problems you actually care about.
You don't get good at agentic coding by memorizing pattern names. You get good by shipping small things, seeing contrast (the lazy prompt beside the structured one), and accumulating taste. This site is built for that — and it grows when you add the problem you wish someone had taught you on.
text equivalent · accessibility
- ░ vibe coding — vibe-prompting, runaway loops, the "just one more retry".
- ▓ spec-driven — specs, instruction files, bounded context.
- █ agentic engineering — structured prompts, custom agents, the spec that ships.
A spectrum learners move along — not a destination they arrive at.
Three content kinds, one filter that runs across all of them.
| Kind | The question it answers | Shape |
|---|---|---|
| Path | Teach me a coherent body of work. | Foundation · Agentic · Terminal · Community |
| Recipe | Show me one thing applied to one problem. | 5–15 min, audience-tagged |
| Project Shape | What problem do I want to apply this to? | Iris Lab · Deal Dashboard · Palette Forge · yours next |
Filter everything by audience — researcher · developer · business · creative — and the catalog reshapes around what you build. No pattern-of-the-month. No taxonomy gatekeeping.
Read the north star:
docs/philosophy.md.
Every change to this curriculum runs through the same AI-assisted flow — the repo eats its own dogfood.
flowchart LR
A([💡 Idea or fix]) --> B[/slash-prompt<br/>in Copilot Chat/]
B --> C{{Agent scaffolds<br/>the change}}
C --> D[Open PR]
D --> E[["@reviewer<br/>@docs-auditor"]]
E -->|approved| F[Merge to main]
F --> G((Site rebuilds<br/>& deploys))
G -.next learner.-> A
classDef accent fill:#ff5500,stroke:#ff5500,color:#fff;
classDef ink fill:#0a0a0a,stroke:#0a0a0a,color:#fff;
class B,C accent
class E ink
The slash-prompts:
/add-example·/fix-content·/refresh-content·/propose-topic.
This repo is the contribution platform. The site is the artifact; this repo is the workshop where the curriculum gets shaped — by people learning this in real time, together.
The most valuable thing you can add is the rep you wish you'd had. A recipe that solved a real problem. A project shape from your domain. A diagram that finally made it click. A correction that saves the next person an hour.
╭──────────────────────────────────────────────────────────────╮
│ Got 10 minutes? → Open an issue. Fix a typo. Add a tag. │
│ Got an evening? → Write a recipe. Add a project shape. │
│ Got a weekend? → Propose a module or a community path. │
╰──────────────────────────────────────────────────────────────╯
Run a slash-prompt in Copilot Chat and the matching agent scaffolds the change. Open the PR; @reviewer and @docs-auditor weigh in automatically.
Before you start, the philosophy gives you a 6-question checklist to know if a piece of content is ready: problem · verbs · feedback · audience · capability+version · (optional) name — see docs/philosophy.md §8.
Start here →
.github/CONTRIBUTING.md·.github/AUTHORING.md
A repo built by the community, where AI helps you contribute to a curriculum about contributing with AI. That's the loop.
src/content/ ←── ✦ where most contributions live (markdown)
src/components/ ←── React + interactive diagrams
src/data/ ←── paths, project shapes, content mappings
.github/ ←── the build harness: agents · skills · prompts · instructions
docs/ ←── design spec & north-star philosophy
Two surfaces, two homes: docs/ is what we teach, .github/ is how we edit it. See docs/harness.md for the map.
- Code → MIT · Content & docs → CC BY 4.0 · third-party attributions in
NOTICE. - Security → please don't open public issues for vulnerabilities. See
SECURITY.mdand https://aka.ms/SECURITY.md. - Code of Conduct → Microsoft Open Source Code of Conduct ·
CODE_OF_CONDUCT.md. - Trademarks → use of Microsoft or third-party marks must follow Microsoft's Trademark & Brand Guidelines.
© Microsoft Corporation.
─────────────────────────────────────────────────────────────
The future isn't AI replacing developers.
It's developers — together — figuring out what good looks like.
─────────────────────────────────────────────────────────────
built in the open · shaped by contributors · ⭐ if it helped, so the next person finds it